Structure and ContentOriented Relevance Feedback

1
Structure and Content-Oriented Relevance Feedback

• Lobna Hlaoua
• Karen Sauvagnat Mohand Boughanem
• IRIT (Institut de Recherche en Informatique de
  Toulouse)
• Equipe SIG-RI (Systèmes dInformations
  Généralisées)
• 118, route de Narbonne - 31062 Toulouse cedex 04

2
Outline

• Context
• Structure-Oriented Relevance Feedback
• Motivation
• Algorithm
• Content-Oriented Relevance Feedback
• Combined approaches
• Experiments
• Conclusion and Outlook

3
Context
Content-Oriented RF
Two approaches are applied
Structure-Oriented RF
Which terms and structures can be extracted from
relevant elements?
How using these extracted items to enrich the
query?
4
Why adding structures could be useful in CO query
?
The user expresses the query using simple
keywords.
lttitlegt Ontologies lt/titlegt
Relevant information are usually in specific
types of components.
Refine CO query by adding structure constraint
5
Why adding structures could be useful in
VVCAS/COS query ?

• The user does not have an exact knowledge about
  the structure of the document.
• The query has a vague constraint structure
• Example
• Result

Q //articleabout( "data mining")
NR R NR R
/article1/bdy1/sec7/ip11 /article1/bdy1
/sec4/ss12/ss26 /article1 /article1/bdy
1/sec4/ss12
6
Our contribution

• Adding structural constraints to CO and VVCAS
  queries
• The structure constraint, is a generative
  structure that is shared by the greatest amount
  of relevant elements.

Example Let us consider eri, erk erl erm four
relevant elements ? Er (a set of relevant
elements) having structures si,,s l,sk. and
scores wi , wk , wl respectively.
si, /article/bdy/sec/ss1 sk, /article/bdy/sec/ss1/
ss2 sl, /article/bdy
We suppose wi wk wl .
7
The Extraction of the Generative Structure
si
sk
sl
Common Structure (CS)
wi wk /2
/article/bdy/sec/ss1
wl wi /4
wk /8
/article/bdy
The generative structure is /article/bdy/sec/ss1
8
Adding the generative structure to the query
(CO/VVCAS)
The generative structure is the top ranked
Structure in the set of Common Structure CS
The structure will be added in two formscomplex
form or simplified tag form. Original query
terms are then added to the structural
constraint. Example is the generative
/article/bdy/sec/ss1 CO query data mining".
Results "/article/bdy/sec/ss1(about(.,informat
ion retrieval)" (complex form) "ss1(about(.,"inf
ormation retrieval")" (simple form).
9
Content-Oriented Relevance Feedback
This approach is based on Rocchio algorithm.
Objective select the more expressive words Lets
consider Erer1, er2, ..., erk,... erm , erk
lnk1,..., lnkj,.lnkn and lnkj tjj
we assign a score to terms (tij ) according to
the following formula
For each relevant element (erk), we compute the
score of terms for each term, we sum its scores
in different leaf nodes.
Best term terms are selected according to the
scores in the set of relevant elements Er .
10
Combined Approaches
The combination consists in applied the two
techniques to extend the initial query by adding
the generative structure and the expressive terms.
CO query we add the structure to the set of
terms in the original query with the top ranked
terms extracted from Er .
VVCAS or COS query we added the top ranked
elements to generative structure in the simple
form. The set is added to initial query with the
Boolean operator OR.
New COS/CAS query Q //articleabout( "data
mining") OR //ss1about(k top ranked terms)
11
Experiments
RI Relative Improvement (MAP(RF-run)-
MAP(base-run))/ MAP(base-run)
S-RF Structure-Oriented Relevance Feedback
CS-FB Content and Structure-Oriented Relevance
Feedback t simple form e1 exhaustivitygt1 e2
exhaustivity2 15 top ranked terms.
12
Conclusion Prospects
We presented a new approach for the Relevance
Feedback in XML Retrieval, which consists in
enriching query by structural constraint
extracted from elements judged relevant.
Future work generating a good generative
structure, by considering only specific
tags using the context of relevant elements in
Relevance Feedback.
13
Thank You
14
Algorithm
For each eri? Er For each erj ? Er-eri
SCA(eri, erj) function to retrieve the Smallest
Common Ancestor
SCA(eri, erj) If spi.firstspj.first If
spi.lastspj.last, then wi ? wiwj
if ? ep(spp,wp) ?
SC/spp.lastspi.last
then wp ? wp wi If spi.last ? spj.last,
then spj ? tail(spj)
wj ? wj/2
SCA(eri, erj)